An MCP gateway sits between AI applications or agents and the servers that expose tools and data through the Model Context Protocol. Instead of connecting each agent to each server directly, teams route those connections through one gateway. The gateway keeps a registry of approved servers, checks who is calling and which tools that caller may use, applies policies such as rate limits, and logs each request and response. The idea resembles an LLM gateway, which manages calls to language models, while an MCP gateway manages calls to tools.
For example, a company with separate MCP servers for its CRM, ticketing system, and data warehouse can let a support agent read tickets and allow only a finance agent to query revenue tables. Gateway logs also show which tool and which data an agent used in a given run, which helps a team reconstruct how the agent reached a result.